A language model specialized in GeneXus
KBbridge-v3 beats the best frontier model at writing GeneXus — at zero cost, on your own infrastructure.
Even when a frontier model is handed all of KB Sync's format skills on every call — its most favourable scenario — KBbridge-v3 beats it on all four quality metrics: syntax (91.2 vs 88.0), exact parameter signature (84.4 vs 53.4), method validity and domain knowledge (78.7 vs 75.4). And it does so without those skills, because the knowledge lives in the weights: no per-token spend, no context tax, and no sending your data to an external API. This is the model we publish on Hugging Face.
1 · Tokens and money
To get a frontier model to write GeneXus, you have to pay for the skills on every call.
When we say "skills" we don't mean the GeneXus manuals. We mean the skills KB Sync publishes: the format-specification files KB Sync writes into the working folder so that any AI agent learns to read and write the .gxSource export format. Without those skills, no frontier model knows how to write it (≤2.6% valid). The only way to get them there is to inject those ~21,600 tokens into the prompt of every request. KBbridge-v3 doesn't need them: that knowledge is in the weights.
| Model | Input tokens | Actual cost | Where it runs |
|---|---|---|---|
| KBbridge-v3 | 0.2M | US$ 0 | Own GPU |
| GPT-5.5 + skills | 11.2M (56×) | US$ 14.04 | External API |
| GLM-5.2 + skills | 11.0M | US$ 4.18 | External API |
| Opus 4.8 + skills* | 12.0M | US$ 70.67* | External API |
* Opus + skills: partial run (337/520) — the most expensive model in the experiment; see the method note at the bottom.
2 · The structural advantages
Three costs of the frontier approach that KBbridge doesn't carry.
Cost per use
$0 vs per-tokenKBbridge runs on a GPU that is already paid for: zero marginal cost. The frontier model pays for input tokens on every call, and since the format skills are mandatory, that spend isn't optional. At production scale the gap grows linearly, forever.
Context tax
+58× · +2.4× latInjecting the skills multiplies the input ~58× and latency ~2.4× (measured). Those ~21,600 tokens compete with the actual problem context: on large tasks or long sessions the window fills up and the model truncates and degrades. KBbridge leaves the whole window for the problem.
Data privacy
Local vs APIThe frontier approach sends your Knowledge Base code to an external API on every request. KBbridge runs local and offline: the code never leaves the company. In a GMP or regulated context that can be a compliance requirement, not a preference.
Evidence of the context tax: feeding KBbridge the skills in context makes it worse (syntax and signature both drop). It already knows what the skills say — the extra context only gets in the way. For a frontier model they are indispensable; for KBbridge they are dead weight. KBbridge's recommended configuration is without skills in context.
3 · Quality deep-dive
KBbridge-v3 beats frontier-with-skills on all four quality metrics.
Every model evaluated under the same protocol: syntax validated with the official GeneXus ANTLR parser from KB Editor, and a generous max_tokens so nothing gets truncated. The frontier models run with KB Sync's format skills injected (their best-case scenario). KBbridge-v3 still beats them on all four metrics — on syntax, and by ~31 points on exact parameter signature (84.4 vs 53.4), which is what lets the object drop into the KB with no manual editing. The full table:
| Configuration | parse | parm | method | MCQ | input | cost |
|---|---|---|---|---|---|---|
| KBbridge-v3 | 91.2⁑ | 84.4 | 89.6 | 78.7 | 0.2M | $0 |
| GPT-5.5 (no skills) | 3.1 | 11.5 | 94.2† | 76.6 | 0.2M | $6.66 |
| GPT-5.5 + skills | 88.0 | 53.4 | 85.9 | 75.4 | 11.2M | $14.04 |
| GLM-5.2 (no skills) | 2.6 | 4.6 | 88.0 | 69.3 | 0.2M | $1.21 |
| GLM-5.2 + skills | 74.4 | 55.7 | 88.5 | 68.1 | 11.0M | $4.18 |
| Opus 4.8 (no skills) | 0.0 | 0.8 | 90.6 | 72.6 | 0.4M | $12.07 |
| Opus 4.8 + skills* | 74.4 | 55.7 | 88.5 | 78.8* | 12.0M | $70.67* |
| DeepSeek-V4-Flash (no skills) | 0.5 | 22.9 | 91.6† | 63.5 | 0.4M | flash · low |
| DeepSeek-V4-Flash + skills | 63.9 | 42.0 | 94.8† | 65.0 | 12.5M | flash · low |
- →Fine-tuning is the differentiator: a generalist without the skills never gets past ~3% valid syntax; with ~21,600 tokens of skills on every call the best frontier model reaches 88%. KBbridge-v3 reaches 91.2% with no skills — the format knowledge is in the weights, not in the context.
- →Parameter signature — KBbridge-v3 dominates: 84.4% vs 53-56% for frontier-with-skills. Without the exact signature, the object doesn't integrate into the KB without manual editing.
- →Domain knowledge: the skills don't move it (they're format, not concepts). KBbridge leads and knows more GeneXus than any of them.
⁑ v3 codegen n=182: 9 giant objects (gold >10 KB) timed out on the GPU (slow/loaded at run time) and were excluded; v3's parseRate is over the 182 measurable ones (the other models: n=191). It doesn't change the ranking — v3 leads by a wide margin on signature and knowledge, neither of which depends on those 9.
Re-scoring: all models re-scored with the validator's improved grammar (−57% false-positive parse failures), same parser and protocol for everyone.
Method: held-out objects that no model saw during training; syntax validated with the official GeneXus ANTLR parser from KB Editor, normalized against the real object.
About "method" (method validity) — do not use it to rank (†): the validator's method database still lacks fundamental GeneXus methods (.Call, .Udp, .AddItem), so it penalizes valid code and, worse, rewards models that write less code (which is why base Qwen scores 93-98% despite being useless on parse/parm).
* Opus 4.8 + skills: full codegen (191/191); partial MCQ (146/329) — credits ran out.
4 · Status and what's next
KBbridge-v3: frontier quality, with every advantage — and it gets better with your own data.
KBbridge-v3 (Qwen 3.8-27B base, LoRA over real GeneXus Knowledge Bases) beats the best frontier-with-skills on all four quality metrics — at zero cost, local, with no context tax. And it isn't a ceiling: the data flywheel (the GeneXus compiler as a verifier) lets us keep raising quality with your own data, without the KB code ever leaving the company.
- →Generalization: measuring v3 on Knowledge Bases it has never seen — the evaluation that matters most for adopting it on your own KB.
- →Data flywheel: the GeneXus compiler as a verifier (trajectory-level rejection sampling) to keep raising quality in the next versions.
- →Per-customer cartridge: the model can specialize on each customer's KB, without their data mixing with anyone else's or leaving their infrastructure.
5 · Download the model
Get KBbridge-v3 and run it wherever you want.
We publish the model on Hugging Face in two formats. Pick the one that matches who will use it: vLLM if you're serving a whole team, GGUF if you're running it on your own machine.
A GPU server that answers the whole team at once. This is the production format: maximum throughput, 256K of context and speculative decoding. It asks for a datacenter-class GPU (Hopper/Blackwell).
To run it on your own laptop or desktop with LM Studio, Ollama or llama.cpp. No datacenter GPU needed, and it works offline.
Frequently asked questions
What people ask us most about KBbridge A.I
What is KBbridge-v3?
KBbridge-v3 is a language model specialized in GeneXus: a Qwen 3.8-27B base fine-tuned with LoRA over real GeneXus Knowledge Bases, served in FP8 with 256K of context on our own GPU. It knows how to write the .gxSource export format without any skills in the prompt, because that knowledge ended up in the model's weights.
What are the "skills" that frontier models need?
They are the format-specification files KB Sync publishes into the working folder, so that any AI agent learns to read and write the .gxSource export format. They are not the GeneXus manuals or documentation: they are the rules of the interchange format. They take about 21,600 tokens, and a frontier model needs them in the prompt of every request; KBbridge-v3 does not, because it has them in its weights.
Why does a frontier model need them on every call?
Because no generalist model learned the .gxSource format during training: without the skills it never gets past 3% valid syntax. The only way to make it write GeneXus is to inject those ~21,600 tokens into the prompt of every request, which multiplies the input ~58× and latency ~2.4×.
How much does KBbridge-v3 cost to run?
Zero per query. The model runs on an already-amortized GPU, so the marginal cost of each request is nil. The same workload (520 objects) with a frontier model plus the skills in context cost between US$ 4 and US$ 70 depending on the model.
Which build should I download: vLLM or GGUF?
vLLM if you're serving several people at once from a datacenter GPU — there you want KBBridge-v3-FP8 (29 GB) for production, or KBBridge-v3 in BF16 (51 GB) if you plan to do your own fine-tuning. GGUF if you're running it alone on your machine with LM Studio, Ollama or llama.cpp: Q4_K_M (16 GB) is the default choice and Q8_0 (28 GB) gives more fidelity if you have the VRAM. They are all at huggingface.co/KBBridge.
Does my Knowledge Base code leave the company?
No. KBbridge-v3 runs local and offline: your KB code never leaves your infrastructure. The frontier approach, by contrast, sends that code to an external API on every request — in regulated or GMP contexts that can be a compliance blocker.
How were these results measured?
With a benchmark of 580 held-out objects per model (191 code generation, 329 multiple choice and 60 data model) that no model saw during training. Syntax was validated with the official GeneXus ANTLR parser from KB Editor, using the same grammar and the same protocol for every model. The frontier models were run with the skills injected — that is, in their best-case scenario.
KBbridge · The bridge between GeneXus and AI CodingKBbridge-v3 · Qwen 3.8-27B fine-tuned with LoRA over real GeneXus Knowledge Bases, served in FP8 (256K of context, the value Qwen recommends for 3.8) on our own GPU. Benchmark of 580 held-out objects per model (191 codegen + 329 MCQ + 60 data model), same protocol for all; the frontier models run with KB Sync's format skills injected (their best-case scenario), via OpenRouter. Syntax validated with the official GeneXus parser from KB Editor (re-scored grammar, the same one for all). Numbers reproducible with the project's evaluation harness.